Datasets for Towards Highly Efficient Semantic Communication via Semantic-aware Compression on Time Series Data

Goal-oriented semantic communication offers a promising solution to the challenges of communication and energy costs raised by increasing Internet of Things (IoT) data traffic. While deep learning-based methods are commonly employed, they face limitations when handling sequential time series data and incur high computational overhead, which renders them unsuitable for resource-constrained IoT environments. While conventional data compression techniques play a crucial role in reducing transmission and storage costs, they are constrained by their focus on data reconstruction quality rather than downstream task accuracy. In this paper, we extend SHRINK, a semantic-aware time-series data compression method, incorporating semantic quantization and transformation to support direct analytics within a goal-oriented edge analytics framework. We first assess SHRINK\u2019s effectiveness through a case study on outlier detection and then further tailor it for this task. We conduct extensive experiments on 193 datasets across multiple domains using three widely used outlier detection methods: Isolation Forest (IForest), Local Outlier Factor (LOF), and SAND. Our results show that direct outlier detection on compressed semantics achieves accuracy comparable to that on raw data, or even higher, in nearly all cases, while accessing only 1\u20135% of the original data volume with up to two orders of magnitude runtime speedup.